What I changed, checkable before we talk.

Proof

50,000 SKU warehouse cutover in Melbourne; order processing time fell about 20%
USD 30M Cross-border e-commerce P&L; about USD 5M of inventory freed
2,708 Public marketplace claims audited; 57% written on or before delivery day
10,756 Passing tests on the deterministic replenishment engine (private repository)

What each claim rests on

I answered for five of these seven risks in live operations, and you have my word for those. Two you can re-run yourself.

The dot is how far I went; the ring is how far you can check it without me. From left to right: not in scope, described, built and tested outside a live system, answered for it in a live operation.

  • The system is not installed, configured and supportedOperations

    How far I went: Built and tested outside a live system (Tested, private repository). An inventory and replenishment engine with 10,756 passing tests, in a private repository.

    What you can check without me: Built and tested outside a live system (Public and re-runnable). The receiving kit is open source, with its tests. See the code on GitHub.

  • The record was already wrongFinance

    How far I went: Answered for it in a live operation (My field work). I ran inventory accuracy and cycle-count programs across a distribution network.

    What you can check without me: Built and tested outside a live system (Public and re-runnable). A re-runnable audit of a public marketplace dataset, and the integrity guards of the open-source kit. See the audit and the code on GitHub.

  • The system is set up for the manual, not the floorOperations

    How far I went: Answered for it in a live operation (My field work). In Melbourne my role was split between the desk and the floor, through the full cutover from the old warehouse system to the new one.

    What you can check without me: Nothing you can run without me.

  • The incentive pays for the wrong thingOperations

    How far I went: Answered for it in a live operation (My field work). On a floor walk in one of my consulting engagements, a per-operator productivity bonus was producing mixed, short and damaged orders. I retired it.

    What you can check without me: Nothing you can run without me.

  • A bad number writes straight to the live systemFinance

    How far I went: Built and tested outside a live system (Tested, private repository). A draft stage, a signed approval that expires, a flag for finance above a threshold and a way back. No client runs it in production yet.

    What you can check without me: Nothing you can run without me: the decision card below is a sample with invented figures.

  • Nobody owns it after go-liveThe chief executive

    How far I went: Answered for it in a live operation (My field work). In the Melbourne cutover I trained 23 operators and stayed on shift until the workarounds died.

    What you can check without me: Nothing you can run without me.

  • The cash and the return never reach the P&LFinance

    How far I went: Answered for it in a live operation (My field work). A USD 30M cross-border P&L, where about USD 5M of inventory was freed.

    What you can check without me: Nothing you can run without me.

Field work is my own account: it illustrates, it does not validate. None of my systems has a client in production yet.

Field work

Thirteen years across operations, planning and consulting. Unarchived field diagnoses were consolidated into The Deployment Chasm and encoded into the checks below.

  • System changeover without floor bypass (AGCO Corporation, Melbourne, 2022 to 2024). Outbound Leader, split between the desk and the floor: picking, packing and the carriers. Through the full cutover of a 50,000-SKU distribution centre with a team of 20, I closed the legacy warehouse system and opened Manhattan Associates, trained 23 operators on RF scanning, and stayed on shift until the workarounds ended. Order processing time fell about 20% (a shift estimate without retained logs).
  • Working capital and demand rules under rapid scale (cross-border e-commerce, 2020 to 2022). Owned a USD 30M P&L spanning direct consumer sales and wholesale. Through 300% annual growth, replaced judgement buffers with aggregate statistical demand forecasting and wrote the operating procedures that kept volume moving, freeing about USD 5M of inventory.
  • Multi-site reporting and replenishment control (Chile and Argentina, 2016 to 2020). Ran a USD 15M budget across three logistics sites and about 100 people, replacing lagging weekly packs with daily operational reporting. Earlier led 15 planners supplying more than 200 retail stores (partly remotely from Sydney), building forecasting and replenishment models while automating financial reports.
  • Cost allocation and pre-cutover savings (2013 to 2016). Analysed a USD 100M inventory position with cost-allocation models and cycle-count controls, and identified about USD 2M a year in purchasing and logistics savings alongside floor procedures for an earlier warehouse-system rollout.
  • Cross-industry diagnostics and incentive repair (since 2013). Across more than seven independent consulting engagements in warehousing, retail, e-commerce and manufacturing, I built process maps, KPI dashboards and floor procedures. On one floor walk, tracing mixed, short and damaged orders back to a per-operator productivity bonus showed how local metrics break downstream quality; I retired the bonus metric and measured customer satisfaction through zero complaints and response time instead.

The Melbourne employer is named once, here. Earlier employers and consulting clients stay unnamed; for proof you can verify without calling anyone, start with the public audit and tested tools below.

Published audit you can re-run

57% of 2,708 “not received” claims on delivered orders were written on or before the recorded delivery day
1,542 late orders 73 at flagged sellers 1,093 with no clear cause

When “not received” means late. A forensic triage on a public Brazilian marketplace dataset showing which record lied before anyone blames the warehouse or the carrier. The Python code and its tests are public and re-runnable.

Read the audit

Built systems

I build diagnostic and planning tools so I do not arrive on a floor with a blank spreadsheet. None is in commercial production at a client yet; they show how operational physics becomes tested software before touching a live system.

  • Inventory and replenishment engine. 10,756 passing tests in Python and none failing, clean install on 19 November 2026 (private repository). Decision cards set the engine’s recommendation beside the operator’s play and the status quo, staging changes behind HMAC-SHA256-signed approval and rollback.
  • Receiving kit. Open source under the MIT licence, with 1,141 passing tests and none failing, clean install on 18 November 2026.
  • Field kit. 1,187 passing tests on 20 October 2026. Checks an event log against the five-column contract (case, activity, timestamp, who, where) before analysis runs; integrity guards catch one-to-many join fanout, unbalanced blocked stock and negative stock against planted defects.
  • Weightlifting planning application. TypeScript monorepo with about 1,332 unit tests, 31 integration tests on embedded PostgreSQL and an adversarial review agent. Building it left a builder-side adoption postmortem: putting a number on what an expert used to judge by eye can still trigger refusal.

The rule I build to:

  1. Check the extract before the math. Run integrity guards on raw tables: one-to-many join fanout, unbalanced blocked stock, negative stock and double-counted stock in transit. If a number fails, nothing is written to the live system.
  2. Stage three plays in draft. Put the engine’s recommendation, the operator’s own adjustment, and the status quo side by side in draft with their working-capital and service-level trade-offs.
  3. Sign off or override with a reason. The buyer or supervisor approves, modifies, or rejects the draft card; the decision and its reason are logged before execution.
  4. Keep the way back. Every approved change carries a rollback path and a 90-day check against the floor baseline.

One decision card, as it waits for a signature. Pick a play to see what the sign-off records, then take the lead time away to see what the engine does with a cost it cannot measure.

Decision card: one purchase order for one stock item Draft: waiting for sign-off

A synthetic sample of the instrument, not a client’s data. The stock, sales and order figures are invented; the three plays were measured on them by the engine’s own purchase rule.

First, the integrity gate

Two probes held the extract before any arithmetic ran:

  • A one-to-many join. Joining the open purchase order to its delivery schedule repeated one order line three times, so the order on the way would have counted three times over.
  • Blocked stock that does not balance. On one order line, more had been released from blocked stock and returned to the supplier than was ever received into it.

Both were fixed at the source and the probes ran clean. Only then were the plays measured.

Then, three plays on one yardstick

USD 30,000 of stock on hand, a supplier lead time of 30 days, a cover target of at most 45 days, and a buyer who wants to order USD 48,000. Margin at risk counts double against capital in the ranking.

Without the lead time, margin at risk cannot be measured, so every play says so instead of printing zero. Capital leaves the ranking too: on its own it would always reward the smallest order for what is not known.

Last, the sign-off record

Picked: play A. The buyer approves the engine’s play. The record keeps the play, who signed, when, and why.

Picked: play B. An override, logged with its reason: USD 24,000 more capital buys twelve more days of cover and clears the margin at risk.

Picked: play C. Doing nothing is a decision too. It is logged with its reason and its cost: USD 11,096 of margin at risknot measured, for want of a lead time.

  • Who signs. The buyer. Above a financial threshold, set on the server and not by whoever signs, the card goes to finance instead.
  • Before the write. The approval is signed and expires, and the write is refused if the spreadsheet changed after the plan was made.
  • The way back. A backup is taken before the write, and one undo puts the sheet back as it was.
  • Where it writes. To a working copy of the spreadsheet, never to the live system. The original stays as it came in, beside a twin workbook that carries the decision.

The code is on GitHub.

Writing

The Deployment Chasm, a manuscript on why enterprise and supply-chain systems fail between the contract and the floor. Leaf through the sample as a book or read it full screen.

Open sample book

Ask for a walk-through